From e6135fbeca4bbedd24b5f8c55fb765a97bdc78ad Mon Sep 17 00:00:00 2001 From: Michael Vrable Date: Tue, 3 Aug 2010 14:10:21 -0700 Subject: [PATCH] Improve the reading back of objects committed to the journal. Implement a cache of memory-mapped log files so that when multiple objects are requested we can re-use the mapping. Make log files fixed sizes (call ftruncate when opening the log file) so the entire thing can be memory mapped at the start. --- bluesky/bluesky-private.h | 7 +- bluesky/file.c | 1 + bluesky/log.c | 197 +++++++++++++++++++++----------------- bluesky/util.c | 2 + 4 files changed, 117 insertions(+), 90 deletions(-) diff --git a/bluesky/bluesky-private.h b/bluesky/bluesky-private.h index ad8dea1..2932417 100644 --- a/bluesky/bluesky-private.h +++ b/bluesky/bluesky-private.h @@ -241,8 +241,13 @@ uint32_t crc32c_finalize(uint32_t crc); struct _BlueSkyLog { char *log_directory; GAsyncQueue *queue; - int fd; + int fd, dirfd; int seq_num; + GSList *committed; + + /* Cache of log segments which have been memory-mapped. */ + GMutex *mmap_lock; + GHashTable *mmap_cache; }; BlueSkyLog *bluesky_log_new(const char *log_directory); diff --git a/bluesky/file.c b/bluesky/file.c index cb9d5f0..34e103f 100644 --- a/bluesky/file.c +++ b/bluesky/file.c @@ -231,6 +231,7 @@ void bluesky_block_fetch(BlueSkyInode *inode, BlueSkyBlock *block, bluesky_string_ref(block->data); g_mutex_unlock(block->cloudref->lock); block->type = BLUESKY_BLOCK_CACHED; + g_atomic_int_add(&inode->fs->cache_total, 1); } /* Write the given block to cloud-backed storage and mark it clean. */ diff --git a/bluesky/log.c b/bluesky/log.c index 152f598..9b4b62b 100644 --- a/bluesky/log.c +++ b/bluesky/log.c @@ -66,6 +66,61 @@ static void writebuf(int fd, const char *buf, size_t len) } } +static void log_commit(BlueSkyLog *log) +{ + int batchsize = 0; + + if (log->fd < 0) + return; + + fdatasync(log->fd); + while (log->committed != NULL) { + BlueSkyCloudLog *item = (BlueSkyCloudLog *)log->committed->data; + g_mutex_lock(item->lock); + item->pending_write &= ~CLOUDLOG_JOURNAL; + item->location_flags |= CLOUDLOG_JOURNAL; + g_cond_signal(item->cond); + g_mutex_unlock(item->lock); + log->committed = g_slist_delete_link(log->committed, log->committed); + batchsize++; + } + + if (bluesky_verbose && batchsize > 1) + g_print("Log batch size: %d\n", batchsize); +} + +static gboolean log_open(BlueSkyLog *log) +{ + char logname[64]; + + if (log->fd >= 0) { + log_commit(log); + close(log->fd); + log->seq_num++; + log->fd = -1; + } + + while (log->fd < 0) { + g_snprintf(logname, sizeof(logname), "log-%08d", log->seq_num); + log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600); + if (log->fd < 0 && errno == EEXIST) { + fprintf(stderr, "Log file %s already exists...\n", logname); + log->seq_num++; + continue; + } else if (log->fd < 0) { + fprintf(stderr, "Error opening logfile %s: %m\n", logname); + return FALSE; + } + } + + if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) { + fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname); + } + fsync(log->fd); + fsync(log->dirfd); + return TRUE; +} + /* All log writes (at least for a single log) are made by one thread, so we * don't need to worry about concurrent access to the log file. Log items to * write are pulled off a queue (and so may be posted by any thread). @@ -82,33 +137,11 @@ static gpointer log_thread(gpointer d) { BlueSkyLog *log = (BlueSkyLog *)d; - /* If there are multiple log items to write, we may write more than one - * before calling fsync(). The committed list is used to track all the - * items that should be marked as committed once that final fsync() is - * done. */ - GSList *committed = NULL; - - int dirfd = open(log->log_directory, O_DIRECTORY); - if (dirfd < 0) { - fprintf(stderr, "Unable to open logging directory: %m\n"); - return NULL; - } - while (TRUE) { if (log->fd < 0) { - char logfile[64]; - g_snprintf(logfile, sizeof(logfile), "log-%08d", log->seq_num); - log->fd = openat(dirfd, logfile, O_CREAT|O_WRONLY|O_EXCL, 0600); - if (log->fd < 0 && errno == EEXIST) { - fprintf(stderr, "Log file %s already exists...\n", logfile); - log->seq_num++; - continue; - } else if (log->fd < 0) { - fprintf(stderr, "Error opening logfile %s: %m\n", logfile); + if (!log_open(log)) { return NULL; } - fsync(log->fd); - fsync(dirfd); } BlueSkyCloudLog *item @@ -116,6 +149,7 @@ static gpointer log_thread(gpointer d) g_mutex_lock(item->lock); g_assert(item->data != NULL); + /* The item may have already been written to the journal... */ if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) { g_mutex_unlock(item->lock); bluesky_cloudlog_unref(item); @@ -124,12 +158,23 @@ static gpointer log_thread(gpointer d) item->pending_write |= CLOUDLOG_JOURNAL; - off_t logsize = lseek(log->fd, 0, SEEK_CUR); struct log_header header; struct log_footer footer; + size_t size = sizeof(header) + sizeof(footer) + item->data->len; + off_t offset = 0; + if (log->fd >= 0) + offset = lseek(log->fd, 0, SEEK_CUR); + + /* Check whether the item would overflow the allocated journal size. + * If so, start a new log segment. We only allow oversized log + * segments if they contain a single log entry. */ + if (offset + size >= LOG_SEGMENT_SIZE && offset > 0) { + log_open(log); + offset = 0; + } header.magic = GUINT32_TO_LE(HEADER_MAGIC); - header.offset = GUINT64_TO_LE(logsize); + header.offset = GUINT64_TO_LE(offset); header.size = GUINT32_TO_LE(item->data->len); header.id = item->id; footer.magic = GUINT32_TO_LE(FOOTER_MAGIC); @@ -148,42 +193,18 @@ static gpointer log_thread(gpointer d) writebuf(log->fd, (const char *)&footer, sizeof(footer)); item->log_seq = log->seq_num; - item->log_offset = logsize + sizeof(header); + item->log_offset = offset + sizeof(header); item->log_size = item->data->len; - logsize += sizeof(header) + sizeof(footer) + item->data->len; + offset += sizeof(header) + sizeof(footer) + item->data->len; - committed = g_slist_prepend(committed, item); + log->committed = g_slist_prepend(log->committed, item); g_mutex_unlock(item->lock); - /* Force an fsync either if we will be closing this log segment and - * opening a new file, or if there are no other log items currently - * waiting to be written. */ - - if (logsize >= LOG_SEGMENT_SIZE - || g_async_queue_length(log->queue) <= 0) - { - int batchsize = 0; - fdatasync(log->fd); - while (committed != NULL) { - item = (BlueSkyCloudLog *)committed->data; - g_mutex_lock(item->lock); - item->pending_write &= ~CLOUDLOG_JOURNAL; - item->location_flags |= CLOUDLOG_JOURNAL; - g_cond_signal(item->cond); - g_mutex_unlock(item->lock); - committed = g_slist_delete_link(committed, committed); - batchsize++; - } - /* if (batchsize > 1) - g_print("Log batch size: %d\n", batchsize); */ - } - - if (logsize < 0 || logsize >= LOG_SEGMENT_SIZE) { - close(log->fd); - log->fd = -1; - log->seq_num++; - } + /* Force an if there are no other log items currently waiting to be + * written. */ + if (g_async_queue_length(log->queue) <= 0) + log_commit(log); } return NULL; @@ -197,6 +218,14 @@ BlueSkyLog *bluesky_log_new(const char *log_directory) log->fd = -1; log->seq_num = 0; log->queue = g_async_queue_new(); + log->mmap_lock = g_mutex_new(); + log->mmap_cache = g_hash_table_new(NULL, NULL); + + log->dirfd = open(log->log_directory, O_DIRECTORY); + if (log->dirfd < 0) { + fprintf(stderr, "Unable to open logging directory: %m\n"); + return NULL; + } g_thread_create(log_thread, log, FALSE, NULL); @@ -225,8 +254,7 @@ void bluesky_log_finish_all(GList *log_items) } /* Memory-map the given log object into memory (read-only) and return a pointer - * to it. FIXME: Use some type of cache, map entire log segments, and use - * reference counting? */ + * to it. */ static int page_size = 0; BlueSkyRCStr *bluesky_log_map_object(BlueSkyLog *log, @@ -236,44 +264,35 @@ BlueSkyRCStr *bluesky_log_map_object(BlueSkyLog *log, page_size = getpagesize(); } - int dirfd = open(log->log_directory, O_DIRECTORY); - if (dirfd < 0) { - fprintf(stderr, "Unable to open logging directory: %m\n"); - return NULL; - } + BlueSkyMmap *map; + g_mutex_lock(log->mmap_lock); + map = g_hash_table_lookup(log->mmap_cache, GINT_TO_POINTER(log_seq)); - char logfile[64]; - g_snprintf(logfile, sizeof(logfile), "log-%08d", log_seq); - int fd = openat(dirfd, logfile, O_RDONLY); - close(dirfd); + if (map == NULL) { + char logname[64]; + g_snprintf(logname, sizeof(logname), "log-%08d", log_seq); + int fd = openat(log->dirfd, logname, O_RDONLY); - if (fd < 0) { - fprintf(stderr, "Error opening logfile %s: %m\n", logfile); - return NULL; - } + if (fd < 0) { + fprintf(stderr, "Error opening logfile %s: %m\n", logname); + g_mutex_unlock(log->mmap_lock); + return NULL; + } + + map = g_new0(BlueSkyMmap, 1); - off_t off_start, off_end; - off_start = log_offset; - off_end = off_start + log_size; - off_start &= ~(page_size - 1); - off_end = (off_end + (page_size - 1)) & (page_size - 1); + off_t length = lseek(fd, 0, SEEK_END); + map->addr = (const char *)mmap(NULL, length, PROT_READ, MAP_SHARED, + fd, 0); + map->len = length; + g_atomic_int_set(&map->refcount, 1); - const char *ptr = (const char *)mmap(NULL, off_end - off_start, PROT_READ, - MAP_SHARED, fd, off_start); + g_hash_table_insert(log->mmap_cache, GINT_TO_POINTER(log_seq), map); - if (ptr == NULL) { - fprintf(stderr, "Error mapping logfile: %m\n"); close(fd); - return NULL; } - close(fd); - - BlueSkyMmap *mmap = g_new0(BlueSkyMmap, 1); - mmap->addr = ptr; - mmap->len = off_end - off_start; - g_atomic_int_set(&mmap->refcount, 1); + g_mutex_unlock(log->mmap_lock); - return bluesky_string_new_from_mmap(mmap, - log_offset - off_start, log_size); + return bluesky_string_new_from_mmap(map, log_offset, log_size); } diff --git a/bluesky/util.c b/bluesky/util.c index c4bac19..0bc1c6c 100644 --- a/bluesky/util.c +++ b/bluesky/util.c @@ -89,6 +89,8 @@ BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s) BlueSkyRCStr *bluesky_string_new_from_mmap(BlueSkyMmap *mmap, int offset, gsize len) { + g_assert(offset + len < mmap->len); + BlueSkyRCStr *string = g_new(BlueSkyRCStr, 1); string->mmap = mmap; g_atomic_int_inc(&mmap->refcount); -- 2.20.1